Skip to content

Hi — Gnostic Relevance Activation

Gnostic relevance activation measuring feature agreement with concept center z0 via tanh(2θ) with learnable scale S. Value range: [-1, 1].


Overview

High values indicate strong alignment with the learned manifold; paired with Hj for relevance-irrelevance analysis.

The centered coordinate is θ = (x - z0) / S.


Key Features

  • Learnable z0, S
  • Relevance / agreement measure
  • Custom autograd kernel

Parameters

Parameter Type Default Description
learnable_S bool True Optimize scale
learnable_z0 bool True Optimize center
initial_S float 1.0 Initial scale
initial_z0 float \| None None Initial center
name str \| None None Name
verbose bool False Logging

Attributes

  • S: Stored raw scale parameter tensor
  • z0: Learnable center tensor
  • theta: Most recent centered deviation (θ)
  • last_output: Most recent relevance response

Methods

forward(x, training=True)

Return the activation applied to input.

  • x: Tensor or array-like — Input values of any shape
  • Returns: Tensor with same shape as x

Example Usage

import numpy as np
from machinegnostics.magnet import Hi

activation = Hi()
result = activation(np.array([[0.0]]))
print(result.data.tolist())  # [[0.0]]

Author: Nirmal Parmar